From Distributed Processing Systems to the buzz word of the day and back Miron Livny Wisconsin...

From Distributed Processing Systems to the buzz word of

the day and back

Miron LivnyWisconsin Institutes for Discovery

Madison-Wisconsin

While I do not know (or understand) your

problem, I am sure that my XYZ software will

solve it!

The words of Koheleth son of David, king in Jerusalem ~ 200 A.D.

Only that shall happen Which has happened,Only that occurWhich has occurred;There is nothing newBeneath the sun!

Ecclesiastes Chapter 1 verse 9

Ecclesiastes, (ת ֶל� ,Kohelet ,ֹק�ֶה�"son of David, and king in Jerusalem" alias Solomon, Wood engravingGustave Doré (1832–1883)

Perspectives on Grid ComputingUwe Schwiegelshohn Rosa M. Badia Marian Bubak Marco DaneluttoSchahram Dustdar Fabrizio Gagliardi Alfred Geiger Ladislav Hluchy

Dieter Kranzlmüller Erwin Laure Thierry Priol Alexander ReinefeldMichael Resch Andreas Reuter Otto Rienhoff Thomas Rüter Peter Sloot Domenico

Talia Klaus Ullmann Ramin Yahyapour Gabriele von Voigt

We should not waste our time in redefining terms or key technologies: clusters, Grids, Clouds... What is in a name? Ian Foster recently quoted Miron Livny saying: "I was doing Cloud computing way before people called it Grid computing", referring to the ground breaking Condor technology. It is the Grid scientific paradigm that counts!

6

HTCondor Team 2013

Project Established 1985

7

1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 1978 1983 1985 1992 1993

“Over the last 15 years, Condor has evolved from a concept to an essential component of U.S. and international cyberinfrastructure supporting a wide range of research, education, and outreach communities. The Condor team is among the top two or three cyberinfrastructure development teams in the country. In spite of their success, this proposal shows them to be committed to rapid development of new capabilities to assure that Condor remains a competitive offering. Within the NSF portfolio of computational and data-intensive cyberinfrastructure offerings, the High Throughput Computing Condor software system ranks with the NSF High Performance Computing centers in importance for supporting NSF researchers.”

An anonymous NSF review (04/2013)

HTCONDORWEEK 2014 9

RACF Overview Main HTCondor pools

◦ PHENIX—12.2kCPU◦ STAR—12.0kCPU◦ ATLAS—13.1kCPU

STAR/PHENIX are RHIC detectors◦ Loose federation of individual users

ATLAS—tightly controlled, subordinate to PANDA workflow management, strict structure

Smaller Experiments◦ LBNE◦ Dayabay◦ LSST

Migration to HTCondor

• Torque/Maui had been used for many years– Many issues– Severity & number of problems increased as size of farm increased

• Migration2012 Aug Started evaluating alternatives to Torque/Maui

(LSF, Grid Engine, Torque 4, HTCondor, SLURM)

2013 Jun Began testing HTCondor with ATLAS & CMS

2013 Aug Choice of HTCondor approved by management

2013 Sep HTCondor declared production service

Moved 50% of pledged CPU resources to HTCondor

2013 Nov Migrated remaining resources to HTCondor

10

Services (2)

Batch:• SLC6 migration: SLC5 CEs decommissioned, no

grid job submission to SLC5– SLC5 WNs final migration ongoing

• Batch system migration, from LSF to HTCondor– Goals: scalability, dynamism, dispatch rate, query scaling– Replacement candidates:

• SLURM feels too young• HTCondor mature and promising• Son of Grid Engine fast, a bit rough

– More details of selection process: https://indico.cern.ch/event/247864/session/5/contribution/22/material/slides/0.pdf

11

https://indico.cern.ch/event/247864/session/5/contribution/22/material/slides/0.pdf



It is all (mainly) about automation

“Do what I told you to do and let me know when you are done!”Automation requires dependable, capable and affordable mechanisms that are controlled by software that implements the policies of the end user. This is why mechanisms determine what a computing system can do for you and it is much harder to change/improve mechanisms then policies. We prefer to talk and do policies. However, it is critical that we talk and understand mechanisms. This requires an understanding of principals and problems not solutions.

Computer Science problems do not die

The good (bad) news about these problems is that they have many non-optimal solutions as they are all based on tradeoffs that cannot be (easily) quantified and are affected by frequent changes in how computers are used and the technologies (hardware and software) that is used to implement/build them.

Problems that do not go way

• Name spaces – when you say “give me foo” how do I know what “foo” is?

• Caching – what should I remove when the cache is full? • Co-scheduling – which resource should I hold while I am

waiting for to other resource to become available? • Error propagation – how do I tell you that I failed? • Verification of a computing system – How do I know that

the system does what it is supposed to do? • Checkpoint restart of applications – How much would it

cost to allocate the resource you are using now to someone else?

Multiprogramming

Maximize throughput by overlapping CPU and I/O, can lead to thrashing due to virtual memory contention (poor locality of reference) and therefore requires inter-application protection.

Maximize utilization of all resources. Individual running time is not the focus (objective).

What is fair?

• How long should I wait? • How much should I pay? • How fast will I run? • How predictive will the service be? • How will I know that I am treated fairly? • How will I know that my resources are

allocated fairly?

In 1978 I fell in love with the problem of load

balancing in distributed systems

The paradigm shift of 70’s – computing

hardware sold in small units.

Claims for “benefits” provided by Distributed Processing Systems

– High Availability and Reliability– High System Performance– Ease of Modular and Incremental Growth– Automatic Load and Resource Sharing– Good Response to Temporary Overloads– Easy Expansion in Capacity and/or Function

P.H. Enslow, “What is a Distributed Data Processing System?” Computer, January 1978

Definitional Criteria for a Distributed Processing System

– Multiplicity of resources– Component interconnection– Unity of control – System transparency– Component autonomy

P.H. Enslow and T. G. Saponas “”Distributed and Decentralized Control in Fully Distributed Processing Systems” Technical Report, 1981

Multiplicity of resources

The system should provide a number of assignable resources for any type of service demand. The greater the degree of replication of resources, the better the ability of the system to maintain high reliability and performance

Component interconnection

A Distributed System should include a communication subnet which interconnects the elements of the system. The transfer of information via the subnet should be controlled by a two-party, cooperative protocol (loose coupling).

System transparency

From the users point of view the set of resources that constitutes the Distributed Processing System acts like a “single virtual machine”. When requesting a service the user should not require to be aware of the physical location or the instantaneous load of the various resources

Unity of Control

All the component of the system should be unified in their desire to achieve a common goal. This goal will determine the rules according to which each of these elements will be controlled.

Component autonomy

The components of the system, both the logical and physical, should be autonomous and are thus afforded the ability to refuse a request of service made by another element. However, in order to achieve the system’s goals they have to interact in a cooperative manner and thus adhere to a common set of policies. These policies should be carried out by the control schemes of each element.

Challenges

• Race Conditions…• Name spaces …• Distributed ownership … • Heterogeneity …• Object addressing …• Data caching …• Object Identity … • Trouble shooting …• Circuit breakers …

One centralized queue or many

distributed queues?

Or, the Wait while Idle problem.

BASICS OF A M/M/1 SYSTEM

l

m

Expected # of customers is 1/(1-r), where ( = r/ ) l m is the utilization

When utilization is 80%,you wait on the average 4 units

for every unit of service

M/M/2 – One queue 2 servers

Number Of

CustomersIn the Queue

Utilization

What about 2*M/M/1 ?

l

m

l

m

When utilization is 80%,you wait on the average 4 units

for every unit of service

When utilization is 80%, 25% of the time a customer is

waiting for service while a server is idle

Wait while Idle (WwI)in m*M/M/1

0

m = 2

m = 5

m = 10

m = 20

Prob(WwI)

1

0 1Utilization

In 1983 I wrote a Ph.D. thesis –

“Study of Load Balancing

Algorithms for Decentralized Distributed Processing Systems”

http://www.cs.wisc.edu/condor/doc/livny-dissertation.pdf

Should I stay and wait or should I

move to another queue?

“ … Since the early days of mankind the primary motivation for the establishment of communities has been the idea that by being part of an organized group the capabilities of an individual are improved. The great progress in the area of inter-computer communication led to the development of means by which stand-alone processing sub-systems can be integrated into multi-computer ‘communities’. … “

Miron Livny, “ Study of Load Balancing Algorithms for Decentralized Distributed Processing Systems.”, Ph.D thesis, July 1983.

In 1984 we introduced the concept of “distributed ownership”, developed our first checkpointing capability and started to implement the first version of Condor leveraging a remote I/O capability (split execution) that was developed at our CS department in 1982.

First version of Condor was installed on 20 DEC2 (desk top) workstations to serve our CS department in 1985.

Submit Locally and run Globally

(Here is the work and here are the

resources I bring to the table)

D. H. J Epema, Miron Livny, R. van Dantzig, X. Evers, and Jim Pruyne, "A Worldwide Flock of Condors : Load Sharing among Workstation Clusters" Journal on Future Generations of Computer Systems, Volume 12, 1996

10

Dubna/Berlin

Amsterdam

3

Warsaw

3

1994 Worldwide Flock of Condors

200

3

Madison

Delft

10

3

Geneva

3010

Use resource and job management “gateways” to connect the Condor pools.

Established a Peer to Peer relationship between the pools to support full local control.

Followed the routing approach of message passing networks to establish a connection between the source (owner of the work) and the destination (resource).

In 1996 I introduced the distinction between High Performance Computing (HPC) and High

Throughput Computing (HTC) in a seminar at the NASA Goddard Flight Center in and a month later at the European Laboratory for Particle Physics (CERN). In June of 1997 HPCWire published an interview on High Throughput Computing.

Why HTC?

For many experimental scientists, scientific progress and quality of research are strongly linked to computing throughput. In other words, they are less concerned about instantaneous computing power. Instead, what matters to them is the amount of computing they can harness over a day, a month or a year --- they measure computing power in units of scenarios per day, wind patterns per week, instructions sets per month, or crystal configurations per year.

High Throughput Computingis a 24-7-365 activity and

therefore requires automation

FLOPY (60*60*24*7*52)*FLOPS

Obstacles to HTC

• Ownership Distribution• Customer Awareness• Size and Uncertainties• Technology Evolution • Physical Distribution

(Sociology)(Education)

(Robustness)(Portability)(Technology)

www.cs.wisc.edu/~miron

Resource Allocation

(resource -> job)

vs.

Work Delegation(job -> resource)



Resource AllocationA limited assignment of the “ownership” of a resourceh Owner is charged for allocation regardless

of actual consumptionh Owner can allocate resource to othersh Owner has the right and means to revoke an

allocationh Allocation is governed by an “agreement”

between the client and the ownerh Allocation is a “lease” h Tree of allocations


“We present some principles that we believe should apply in any compute resource management system. The first, P1, speaks to the need to avoid “resource leaks” of all kinds, as might result, for example, from a monitoring system that consumes a nontrivial number of resources.

P1 - It must be possible to monitor and control all resources consumed by a CE—whether for “computation” or “management.”

Our second principle is a corollary of P1:P2 - A system should incorporate circuit breakers

to protect both the compute resource and clients. For example, negotiating with a CE consumes resources. How do we prevent an eager client from turning into a denial of service attack? “Ian Foster & Miron Livny, "Virtualization and Management of

Compute Resources: Principles and Architecture ", A working document (February 2005)


Garbage collectionis the

cornerstone of

resource allocation


Work Delegation

A limited assignment of the responsibility to perform the workh Delegation involved a definition of

these “responsibilities”h Responsibilities my be further

delegatedh Delegation consumes resourcesh Delegation is a “lease”h Tree of delegations


HTCondor Job Submission Options

› leave_in_queue = <ClassAd Boolean Expression>

› on_exit_remove = <ClassAd Boolean Expression>

› on_exit_hold = <ClassAd Boolean Expression>

› periodic_remove = <ClassAd Boolean Expression>

› periodic_hold = <ClassAd Boolean Expression>

› periodic_release = <ClassAd Boolean Expression>

› noop_job = <ClassAd Boolean Expression>

The Grid Movement

Enable (limited) work delegation and remote (distributed) authorization that is based on a global identity namespace and a (small) group of trusted certificate authorities.

Introduced authentication to Distributed Processing Systems.

www.cs.wisc.edu/~miron 55

The grid promises to fundamentally change the way we think about and use computing. This infrastructure will connect multiple regional and national computational

grids, creating a universal source of pervasive and dependable computing power that supports dramatically new classes of applications. The Grid provides a clear vision of what computational

grids are, why we need them, who will use them, and how they will be programmed.

The Grid: Blueprint for a New Computing InfrastructureEdited by Ian Foster and Carl KesselmanJuly 1998, 701 pages.


“ … We claim that these mechanisms, although originally developed in the context of a cluster of workstations, are also applicable to computational grids. In addition to the required flexibility of services in these grids, a very important concern is that the system be robust enough to run in “production mode” continuously even in the face of component failures. … “

Miron Livny & Rajesh Raman, "High Throughput Resource Management", in “The Grid: Blueprint for a New Computing Infrastructure”.


“ … Grid computing is a partnership between clients and servers. Grid clients have more responsibilities than traditional clients, and must be equipped with powerful mechanisms for dealing with and recovering from failures, whether they occur in the context of remote execution, work management, or data output. When clients are powerful, servers must accommodate them by using careful protocols.… “

Douglas Thain & Miron Livny, "Building Reliable Clients and Servers", in “The Grid: Blueprint for a New Computing Infrastructure”,2nd edition


The Ethernet Protocol

IEEE 802.3 CSMA/CD - A truly distributed (and very effective) access control protocol to a shared service.©Client responsible for access control©Client responsible for error detection©Client responsible for fairness


Client

Server

Master

Worker


minp aijbp(i)p(j)

30

i=1

The NUG30 Quadratic Assignment Problem

(QAP)

30

j=1


NUG30 Personal Grid (06/2000)

Managed by one Linux box at Wisconsin

Flocking: -- the main Condor pool at Wisconsin (500 processors)

-- the Condor pool at Georgia Tech (284 Linux boxes)

-- the Condor pool at UNM (40 processors)

-- the Condor pool at Columbia (16 processors)

-- the Condor pool at Northwestern (12 processors)

-- the Condor pool at NCSA (65 processors)

-- the Condor pool at INFN Italy (54 processors)

Glide-in: -- Origin 2000 (through LSF ) at NCSA. (512 processors)

-- Origin 2000 (through LSF) at Argonne (96 processors)

Hobble-in: -- Chiba City Linux cluster (through PBS) at Argonne (414 processors).


Solution Characteristics.

Scientists 4Workstations 1Wall Clock Time 6:22:04:31Avg. # CPUs 653

Max. # CPUs 1007

Total CPU Time Approx. 11 years

Nodes 11,892,208,412

LAPs 574,254,156,532

Parallel Efficiency 92%


Work

ers

The NUG30 Workforce

Condor crash

ApplicationUpgrade

SystemUpgrade


Being a Master

Customer “delegates” task(s) to the master who is responsible for:h Obtaining allocation of resources h Deploying and managing workers on

allocated resourcesh Delegating work unites to deployed

workersh Receiving and processing resultsh Delivering results to customer


Master must be …

› Persistent – work and results must be safely recorded on non-volatile media

› Resourceful – delegates “DAGs” of work to other masters

› Speculative – takes chances and knows how to recover from failure

› Self aware – knows its own capabilities and limitations› Obedience – manages work according to plan› Reliable – can mange “large” numbers of work items

and resource providers› Portable – can be deployed “on the fly” to act as a “sub

master”


Master should not do …

› Predictions …› Optimal scheduling … › Data mining …› Bidding …› Forecasting …


Dear Master, Never assume that what you know is still true and that what you ordered did actually happen!


Every Communitycan benefit from the

services of

Matchmakers!

eBay is a matchmaker


Why? Because ...

.. someone has to bring together community members who have requests for goods and services with members who offer them.h Both sides are looking for each otherh Both sides have constraintsh Both sides have preferences


Being a Matchmaker

› Symmetric treatment of all parties› Schema “neutral” › Matching policies defined by

parties› “Just in time” decisions › Acts as an “advisor” not “enforcer”› Can be used for “resource

allocation” and “job delegation”

The paradigm shift of 00’s – computing

capacity sold on demand for short time periods.

In other words, computing capacity is assumed to be unbounded as long as you have an unbounded CC.


Miron LivnyCenter for High Throughput Computing

Computer Sciences DepartmentUniversity of Wisconsin-Madison

Condorin the Clouds


The words of Koheleth son of David, king in Jerusalem ….

Only that shall happen Which has happened,Only that occurWhich has occurred;There is nothing newBeneath the sun!

Ecclesiastes Chapter 1 verse 9


An autonomous computing (processing, storage and networking) resources with an interface that supports remote invocation of “jobs” and staging of input/output data.

h Looks and feels like any other grid siteh Likely to have proprietary APIs h Likely to have different cost modelsh Likely to have different SLAsh Likely to have different usage policies

Our view of a cloud


What do we do with clouds?

› Turn VMs into “first class citizens” in the Condor framework

› Interact with users in academia and industry who express interest in using computing resources offered by clouds

› Add EC2+S3 to the (long) list of remote resources Condor can harness (or delegate work to)

› Explore possible enhancements to our matchmaking and workflow technologies to support provisioning of cloud resources (including inter-cloud migration)

› Understand the semantics of the EC2+S3 services, protocols and infrastructure so that we can provide a Condor “overlay” that expend local capabilities to include these resources

› Monitor new cloud “formations”


How can I use Condor?

› As a job manager and resource scheduler for a dedicated collection of rack mounted computers

› As a job manager and resource scheduler for a collection of desk-top computers

› As a job manager and a resource scheduler for a collection of batch/grid/cloud systems

› As a job manager and resource scheduler for all of the above

Everything “looks” and is treated like a job

April 19, 2012, 9:02 a.m. EDT

Cycle Computing Ramps Global 50,000-Core Cluster for Schrodinger Molecular Research Utility Supercomputing Leader Facilitates Massive Cluster for Computational Drug Discovery NEW YORK, NY, Apr 19, 2012 (MARKETWIRE via COMTEX) -- Cycle Computing provisioned a 50,000-core utility supercomputer in the Amazon Web Services (AWS) cloud for Schroedinger and Nimbus Discovery to accelerate lead identification via virtual screening. This milestone -- the largest of its kind -- is Cycle Computing's fifth massive cluster in less than two years on the heels of a 30,000 cluster in October 2011, illustrating Cycle's continued leadership in delivering full-featured and scalable cluster deployments. Cycle Computing revealed the cluster creation during today's opening keynote at the AWS Summit in New York City.

Are EC2 Spot instances a Grid, a Cloud or just a Distributed Processing System where resources come and go at

(local) will?

The Open Science Grid

(OSG)

“The members of the OSG are united by a commitment to promote the adoption and to advance the state of the art of distributed high throughput computing (DHTC) – shared utilization of autonomous resources where all the elements are optimized for maximizing computational throughput.”

The OSG addresses these challenges by following a framework that is based on four underlying principles:

•Resource Diversity•Dependability•Autonomy•Mutual Trust

“This dependability needs to be maintained while the services and their software implementations change to meet new needs and incorporate new technologies”

You may have ONE local submit machine managing 100K jobs

on 10K remote machines

In HTCondor we use a two phase matchmaking process to first

allocate a collection of resources to a requestor and then to select

a task to be delegated for execution within the constraints

of these resources

I am D and I am willing to offer you resources

I am S and am looking for resources

Match!Match!

W3

WiWiWiWi

MM

SchedD StartD

User Code/DAGMan

SchedD

G-app G-app G-app

Local

Remote

HTCondor

MM

Grid CE

HT

Con

dor

StartD StartD StartD

PBSLSF HTCondor

MM C-app

C-app

SSH CondorC

FactoryFront End

OSG

Factory

SchedD

MM

HTCondor

MM

C-app C-app C-app

The OSG GildeIn factory uses the SchedD as a resource provisioning

agent on behalf of the (local) SchedD. It decides when, from

where and for how long to keep an acquired resource.

User Code/DAGMan

SchedD

Local

Remote

HTCondor

MM

EC2HTCondor

MM

C-app

C-app

OpSt Spot

FactoryFront End

OSGCloud

Factory

SchedD

VM VM VM

HT

Con

dor

StartD StartD StartD

MM

C-app C-appC-app

The main challenge is to know in advance how much an

application needs, to monitor how much an application

actually consumes to know how much is available and to

react accordingly

We use networks for control, we use networks to move executables and we use

networks to move data. The application may use networks

internally to its work.

When talking about networks, everyone always thinks of

"bandwidth", but indeed there is much more to consider....

DEFINITIONS

"Intermediary" - Anyone between the client and the server. Could the operating systems on either side, the network cards on either side, routers, switches, NATs, and more. "server" - The side listening for an inbound TCP connection or UDP packet. A given process might be both a server and a client. "client" - The side initiating a TCP connection or sending a UDP packet. A given process might be both a server and a client.

Bandwidth - Can you transmit required data quickly enough to meet your needs? You might be constrained by the physical links as well as any intermediaries. For example, many firewalls process packets more slowly than their network connections otherwise support. Various layers of the system itself may limit bandwidth; security (encryption, decryption, checksums) in particular can easily be expensive. Security can also increase the amount of traffic necessary; authentication of both sides can easily add multiple messages to ultimately send a single small message.

CPU - Can your system make requests quickly enough? Can your system take advantage of multiple CPU cores in a single system to manage load? How much CPU does your security subsystem require; security (encryption, decryption, checksums) can use a lot.

Memory – May be limited by physical RAM and swap, kernel configuration, user-level limits, cgroup limits, or per-process limits. Adding a security subsystem will require more memory. - Process memory - Each network connection requires some memory in your process. Your library providing the networking interface is almost certainly doing memory allocation on your behalf. - Kernel memory - Each network connection, including pending, requires some memory from the kernel.

File descriptors - Each connection requires an FD, and listening for incoming connections is another FD. FDs are a finite resource on the client and server.

Multiple layers can impose limits: per process limits (ulimit), per user limits, per process group limits (by cgroups or similar), configurable system-wide limits (/proc/sys/fs/file-max on Linux), or technical limits (Only 2^32 FDs can be described in a single program on Linux. We believe Linux can only manage 2^32 total FDs).

Ports - Only 65535 are available on a given network interface. In practice, the available number will be much smaller: some will be in use by other processes, some are unavailable to non-root processes, and the system configuration will likely limit the range further (/proc/sys/net/ipv4/ip_local_port_range on Linux controls the ephemeral ports). Closed connections may continue to hold ports for a while to reduce the risk of port-reuse (TCP's TIME_WAIT state). In some cases ports can be shared, but this necessitates adding additional identifying information for each TCP connection or UDP packet. All computers behind a NAT share a finite number of ports for all connections to hosts outside of the NAT.

Firewall state - A firewall/router/NAT has limited resources to manage connections that traverse it. It might have limits on simultaneous connections, simultaneous connections being initiated, bandwidth, RAM, or others. If the intermediary runs out of resources, a wide variety of undesirable things might happen: the intermediary may stop processing anything, blocking all traffic; it may break existing connections; it may reject new connections; it may stop processing firewall rules.

What does HTCondor offer today for

network resoures

The SchedD can manage the allocation of data

transfer connections to users/jobs while

monitoring overall I/O and networking activity

The SharedPort daemon reduces the

number of ports used by an HTCondor machine to one

by reversing TCP connections the Condor Connection Broker (CCB) reduces the number of

(outgoing) Ports used by a SchedD

HTCondor monitors FD consumption and

collects HostName resolution statistics

Know what you need, what you use and what is

available!

Using the power of Directed Acyclic Graphs

(DAGs) to support declarative automation

of interdependent tasks.

Example of a LIGO Inspiral DAG (Workflow)

From: Stuart Anderson <[email protected]> Date: February 28, 2010 11:51:32 PM EST To: Condor-LIGO mailing list <[email protected]> Subject: [CondorLIGO] Largest LIGO workflowPete, Here are some numbers you ask about for LIGO's use of DAGs to manage large data analysis tasks broken down by the largest number of jobs managed in different categories:

1) DAG Instance--one condor_dagman process: 196,862.2) DAG Workflow--launched from a single condor_submit_dag but may include multiple automatic sub- or spliced DAGs: 1,120,659.3) DAG Analysis--multiple instances of condor_submit_dag to analyze a common dataset with results combined into a single coherent scientific result: 6,200,000.4) DAG Total--sum over all instances of condor dagman run: O(100M).

P.S. These are lower bounds as I did not perform an exhaustive survey/search, but they are probably close.

Thanks.

mailto:[email protected]

mailto:[email protected]

Using the power of the “data-flow” model

support planning and enable automation

Homework – Mechanisms to

manage opportunistic

storage.

Date post:	26-Dec-2015
Category:	Documents
Upload:	lillian-russell
View:	213 times
Download:	0 times

From Distributed Processing Systems to the buzz word of the day and back Miron Livny Wisconsin...

Documents